Monitoring and control for NGL recovery plant

The thesis explores the production of natural gas liquids (NGL) and the challenge of monitoring and controlling the fractionation process. NGLs are the C2+ hydrocarbon fraction contained in natural gas, which includes useful feedstocks for industrial production processes. Since NGLs have greater economic value compared to natural gas, their recovery has become increasingly economically significant, leading to a need for efficient fractionation. This energy-intensive process is typically conducted in separation trains that include cryogenic distillation columns. Given the high cost of composition analyzers and the related significant delays, this work proposes the use of only indirect composition control strategies, as well as data-driven control strategies to achieve the desired product quality and optimize the plant energy consumption under typical disturbances. Feedforward neural networks (FFNs) were used for the development of soft sensors used in data-driven control schemes. Given the multitude of data made available by the process simulator, this work aims to develop a demethanizer digital twin that can approximate the column dynamics with reduced computation time. Long Short-Term Memory neural networks (LSTM), along with physical knowledge, were used to develop different neural network architectures compared to select the most suitable for the surrogate model development. Realistic measurement noises were considered to accurately reflect the measurements of real industrial plants and only easy-to-measure variables were used as input data for the developed neural model. Overall, the research presents an energy-efficient NGL recovery offering a cost-effective and efficient alternative to traditional measuring instruments. Moreover, the study illustrates a novel application of LSTM for distillation columns digital twins realization, providing a useful tool for optimization, monitoring and control by employing available plant measurements.